[hts-users:01987] Re: HVite in HTS-SLT-demo
- Subject: [hts-users:01987] Re: HVite in HTS-SLT-demo
- From: Udhyakumar N <udhay@xxxxxxx>
- Date: Sat, 23 May 2009 18:07:07 -0400
- Delivered-to: hts-users@xxxxxxxxxxxxxxx
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=gamma; h=domainkey-signature:mime-version:sender:received:in-reply-to :references:date:x-google-sender-auth:message-id:subject:from:to :content-type:content-transfer-encoding; bh=YeJfhgdI+F6/MHcEGNM7AijaX4sh6+21gNcr+FqMgAo=; b=HU1OzKeeLCvpBdwv6MSunf+I39kiX/ZMpR1nUMaAqlBlESpgxnyNssfY6FzR1i3byf tLK67PPIFc7c3S6TSbbF2y5OhBM6973hL3o1JLLrUIWbE10rSxpdEYs/xceKayd+GGff dnvMPn1ECF2PJTQne4SvZ6/q5XYYdKnJrJUic=
- Domainkey-signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=mime-version:sender:in-reply-to:references:date :x-google-sender-auth:message-id:subject:from:to:content-type :content-transfer-encoding; b=C2RaEoS3o2SH3XQRUrCZaTHdaaRH/Ogc3FtwtwO4f5qb4l1tEDif7bHZd6eqgwrmwq IEmHnzZT+9giif+KrkLiDHxZKo6gD59G5DEiI46GPtHErJwj4b7vdyPOXCSVmsc4nkOq /lyOT9ZdmG+o39EsbWzCiPXDPjUAegSFZZ0V8=
I didnt realize the HTS-SLT-Demo trained HSMMs instead of HMMs. I
tweaked the transition probabilities (as Simon King suggested) and the
HVite alignment works fine now. I dumped all the state-level
alignments to the label files and used them in HMGenS with -s option.
HMGenS crashed with "segmentation fault" error.
I am running HTS on 64-bit fedora 7. Here is the actual command and
output (I omitted all the duration info as the state-alignments in lab
files are used for duration)
----
HMGenS -A -B -s -C configs/syn.cnf -D -T 1 -S
data/test/scp/test_align.scp -t 1500 100 5000 -c 0 -H
models/qst001/ver1/cmp/re_clustered_all.mmf.1mix -M data/test/syn
models/qst001/ver1/cmp/tiedlist
HTK Configuration Parameters[22]
Module/Tool Parameter Value
# OPTKIND NEWTON
# GVWEIGHT 1
# HMMWEIGHT 1
# STEPDEC 0.500000
# STEPINC 1.200000
# STEPINIT 1
# MINEUCNORM 0.010000
# GVEPSILON 0.000100
# MAXGVITER 50
# GVHMMLIST
/people/udhay1/HTS/ARCTIC/SLT-LSP-noGV/models/qst001/ver1/cmp/gv.list
# GVMODELMMF
/people/udhay1/HTS/ARCTIC/SLT-LSP-noGV/models/qst001/ver1/cmp/gv.mmf
# USEGV FALSE
# EMEPSILON 0.000100
# MAXEMITER 20
# WINDIR
/people/udhay1/HTS/ARCTIC/SLT-LSP-noGV/data/win
# WINFN StrVec 2 mgc.win1 mgc.win2 StrVec 2
lf0.win1 lf0.win2
# PDFSTREXT StrVec 2 mgc lf0
# PDFSTRORDER IntVec 2 13 1
# PDFSTRSIZE IntVec 2 1 2
# USEALIGN TRUE
# NATURALWRITEORDER TRUE
# NATURALREADORDER TRUE
HGen ML Generating: Transitions Durations Spaces Mixtures Parameters
HMMSet is SHARED
38765 Logical/27804 Physical Models Loaded, VecSize=28
1 MMF input files
PdfStream [1]:
#streams: 1 (vSize=26)
#order: 13
file ext: mgc
1-th window: /people/udhay1/HTS/ARCTIC/SLT-LSP-noGV/data/win/mgc.win1
2-th window: /people/udhay1/HTS/ARCTIC/SLT-LSP-noGV/data/win/mgc.win2
PdfStream [2]:
#streams: 2 (vSize=2)
#order: 1
file ext: lf0
1-th window: /people/udhay1/HTS/ARCTIC/SLT-LSP-noGV/data/win/lf0.win1
2-th window: /people/udhay1/HTS/ARCTIC/SLT-LSP-noGV/data/win/lf0.win2
Generating Label cmu_us_arctic_slt_a0010.lab
Segmentation fault
On Sat, May 23, 2009 at 4:04 AM, Udhyakumar N <udhay@xxxxxxx> wrote:
> Hi,
>
> I am new to HTS (although I have some exp in using HTK). I
> successfully installed and generated waveforms from the HTS Speaker
> dependent training demo.
>
> I am trying to calculate the MCD (Mel-cepstral distortion) for the
> basic HTS output. So I disabled GV,split the SLT database into train
> and test and used HTS-demo script to train the model with just the
> training subset. For the test subset, I separately generated mgc, lf0
> and the cmp files.
>
> I would like to use the HMM cmp models (trained on the "training"
> subset) to generate state-level alignments for the test set (ignoring
> duration models). The HTK alignment tool HVite requires a dictionary,
> so I made a "hacky" version by copying all the context-dependent
> models twice as
>
> ph1 ph1
> ph2 ph2
>
> where ph1 is a full-context model name. I then ran the HVite command
>
> HVite -A -T 1 -a -m -o SWT -y lab -C configs/trn.cnf -H
> models/qst001/ver1/cmp/re_clustered_all.mmf.1mix -I
> data/test/labels/full.mlf -l data/test/align -S
> data/test/scp/test_cmp.scp data/test/labels/full.dict
> models/qst001/ver1/cmp/tiedlist
>
> I didnt get any error, but the alignment failed with message
>
> No tokens survived to final node of network
>
> I guess the there is a mis-match between model/features/dictionary.
> Can anyone help me resolve this issue?
>
> Thanks,
> Udhay
>
- Follow-Ups
-
- [hts-users:01988] Re: HVite in HTS-SLT-demo, Heiga ZEN (Byung Ha CHUN)
- References
-
- [hts-users:01985] HVite in HTS-SLT-demo, Udhyakumar N